Analysis

Category for analysis updates

Agentic AI and the future of cybersecurity

With the rapid expansion of AI technologies, agentic AI is rapidly moving from experimentation to deployment on a scale larger than ever before. As a result, these systems have been given far greater autonomy to perform tasks with limited human input, much to the delight of enterprise magnates.

Companies such as Microsoft, Google, Anthropic, and OpenAI are increasingly developing agentic AI systems capable of automating vulnerability detection, incident response, code analysis, and other security tasks traditionally handled by human teams.

The appeal of using agentic AI as a first line of defence is palpable, as cybersecurity teams face mounting pressure from the growing volume of attacks. According to the Microsoft Digital Defense Report 2025, the company now detects more than 600 million cyberattacks daily, ranging from ransomware and phishing campaigns to identity attacks. Additionally, the International Monetary Fund has also warned that cyber incidents have more than doubled since the COVID-19 pandemic, potentially triggering institutional failures and incurring enormous financial losses.

To add insult to injury, ransomware groups such as Conti, LockBit, and Salt Typhoon have shown increased activity from 2024 through early 2026, targeting critical infrastructure and global communications, as if aware of the upcoming cybersecurity fortifications and using a limited window of time to incur as much damage as possible.

In such circumstances, fully embracing agentic AI may seem like an ideal answer to the cybersecurity challenges looming on the horizon. Systems capable of autonomously detecting threats, analysing vulnerabilities, and accelerating response times could significantly strengthen cyber resilience.

Yet the same autonomy that makes these systems attractive to defenders could also be exploited by malicious actors. If agentic AI becomes a defining feature of cyber defence, policymakers and companies may soon face a more difficult question: how can they maximise its benefits without creating an entirely new layer of cyber risk?

Why cybersecurity is turning to agentic AI

The growing interest in agentic AI is not simply driven by the rise in cyber threats. It is also a response to the operational limitations of modern security teams, which are often overwhelmed by repetitive tasks that consume time and resources.

Security analysts routinely handle phishing alerts, identity verification requests, vulnerability assessments, patch management, and incident prioritisation — processes that can become difficult to manage at scale. Many of these tasks require speed rather than strategic decision-making, creating a natural opening for AI systems to operate with greater autonomy.

Microsoft has aggressively moved into this space. In March 2025, the company introduced Security Copilot agents designed to autonomously handle phishing triage, data security investigations, and identity management. Rather than replacing human analysts, Microsoft positioned the tools to reduce repetitive workloads and enable security teams to focus on more complex threats.

Google has approached the issue through vulnerability research. Through Project Naptime, the company demonstrated how AI systems could replicate parts of the workflow traditionally handled by human security researchers by identifying vulnerabilities, testing hypotheses, and reproducing findings.

Anthropic introduced another layer of complexity through Claude Mythos, a model built for high-risk cybersecurity tasks. While the company presented the model as a controlled release for defensive purposes, the announcement also highlighted how advanced cyber capabilities are becoming increasingly embedded in frontier AI systems.

Meanwhile, OpenAI has expanded partnerships with cybersecurity organisations and broadened access to specialised tools for defenders, signalling that major AI firms increasingly view cybersecurity as one of the most commercially viable applications for autonomous systems.

Together, these developments show that agentic AI is gradually becoming embedded in the cybersecurity infrastructure. For many companies, the question is no longer whether autonomous systems can support cyber defence, but how much responsibility they should be given.

When agentic AI tools become offensive weapons

The same capabilities that make agentic AI valuable to defenders also make it attractive to malicious actors. Systems designed to identify vulnerabilities, analyse code, automate workflows, and accelerate decision-making can be repurposed for offensive cyber operations.

Anthropic offered one of the clearest examples of that risk when it disclosed that malicious actors had used Claude in cyber campaigns. The company said attackers were not simply using the model for basic assistance, but were integrating it into broader operational workflows. The incident showed how agentic AI can move cyber misuse beyond advice and into execution.

The risk extends beyond large-scale cyber operations. Agentic AI systems could make phishing campaigns more scalable, automate reconnaissance, accelerate vulnerability discovery, and reduce the technical expertise needed to launch certain attacks. Tasks that once required specialist teams could become easier to coordinate through autonomous systems.

Security researchers have repeatedly warned that generative AI is already making social engineering more convincing through realistic phishing emails, cloned voices, and synthetic identities. More autonomous systems could further push those risks by combining content generation with independent action.

The concern is not that agentic AI will replace human hackers. Cybercrime could become faster, cheaper, and more scalable, mirroring the same efficiencies that organisations hope to achieve through AI-powered defence.

The agentic AI governance gap

The governance challenge surrounding agentic AI is no longer theoretical. As autonomous systems gain access to internal networks, cloud infrastructure, code repositories, and sensitive datasets, companies and regulators are being forced to confront risks that existing cybersecurity frameworks were not designed to manage.

Policymakers are starting to respond. In February 2026, the US National Institute of Standards and Technology (NIST) launched its AI Agent Standards Initiative, focused on identity verification and authentication frameworks for AI agents operating across digital environments. The aim is simple but important: organisations need to know which agents can be trusted, what they are allowed to do, and how their actions can be traced.

Governments are also becoming more cautious about deployment risks. In May 2026, the Cybersecurity and Infrastructure Security Agency (CISA) joined cybersecurity agencies from Australia, Canada, New Zealand, and the United Kingdom in issuing guidance on the secure adoption of agentic AI services. The warning was clear: autonomous systems become more dangerous when they are connected to sensitive infrastructure, external tools, and internal permissions.

The private sector is adjusting as well. Companies are increasingly discussing safeguards such as restricted permissions, audit logs, human approval checkpoints, and sandboxed environments to limit the degree of autonomy granted to AI agents.

The questions facing businesses are becoming practical. Should an AI agent be allowed to patch vulnerabilities without approval? Can it disable accounts, quarantine systems, or modify infrastructure independently? Who is held accountable when an autonomous system makes the wrong decision?

Agentic AI may become one of cybersecurity’s most effective defensive tools. Its success, however, will depend on whether governance frameworks evolve quickly enough to keep pace with the technology itself.

How companies are building guardrails around agentic AI

As concerns around autonomous cyber systems grow, companies are increasingly experimenting with safeguards designed to prevent agentic AI from becoming an uncontrolled risk. Rather than granting unrestricted access, many organisations are limiting what AI agents can see, what systems they can interact with, and what actions they can execute without human approval.

Anthropic has restricted access to Claude Mythos over concerns about offensive misuse, while OpenAI has recently expanded its Trusted Access for Cyber programme to provide vetted defenders with broader access to advanced cyber tools. Both approaches reflect a growing consensus that powerful cyber capabilities may require tiered access rather than unrestricted deployment.

The broader industry is moving in a similar direction. CrowdStrike has increasingly integrated AI-driven automation into threat intelligence and incident response workflows while maintaining human oversight for critical decisions. Palo Alto Networks has also expanded its AI-powered security automation tools designed to reduce response times without fully removing human analysts from the decision-making process.

Cloud providers are also becoming more cautious about autonomous access. Amazon Web Services, Google Cloud, and Microsoft Azure have increasingly emphasised zero-trust security models, role-based permissions, and segmented access controls as enterprises deploy more automated tools across sensitive infrastructure.

Meanwhile, sectors such as finance, healthcare, and critical infrastructure remain particularly cautious about fully autonomous deployment due to the potential consequences of false positives, accidental shutdowns, or disruptions to essential services.

As a result, security teams are increasingly discussing safeguards such as audit logs, sandboxed environments, role-based permissions, staged deployments, and human approval checkpoints to balance speed with accountability. For now, many companies seem ready to embrace agentic AI, but without keeping one hand on the emergency brake.

The future of cybersecurity may be agentic

Agentic AI is unlikely to remain a niche experiment for long. The scale of modern cyber threats, combined with the mounting pressure on security teams, means organisations will continue to look for faster and more scalable defensive tools.

That shift could significantly improve cybersecurity resilience. Autonomous systems may help organisations detect threats earlier, reduce response times, address workforce shortages, and manage the growing volume of attacks that human teams increasingly struggle to handle alone.

At the same time, the technology’s long-term success will depend as much on restraint as on innovation. Without clear governance frameworks, operational safeguards, and human oversight, the same tools designed to strengthen cyber defence could introduce entirely new vulnerabilities.

The future of cybersecurity may increasingly belong to agentic AI. Whether that future becomes safer or more volatile may depend on how responsibly governments, companies, and security teams manage the transition.

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!


Why DeepSeek V4 is changing the AI model race

DeepSeek has again placed itself at the centre of the global AI race. After drawing worldwide attention with its R1 reasoning model in early 2025, the Chinese company has recently released DeepSeek V4, a new model designed to compete not only on performance, but also on price, openness and efficiency.

The hype around DeepSeek V4 is not based on a single feature. The model comes with a 1 million-token context window, open weights, two versions for different use cases and a strong focus on agentic workflows such as coding, research, document analysis and long-running tasks. In a market still dominated by expensive closed models, DeepSeek is trying to prove that powerful AI does not need to remain locked behind trademarked systems.

A model built for long memory

The most immediate difference between DeepSeek V4 and other models is context length. Both DeepSeek-V4-Pro and DeepSeek-V4-Flash support a 1-million-token context window, meaning they can process inputs far longer than those of older generations of mainstream models. According to DeepSeek’s official release, one million tokens is now the default across all official DeepSeek services.

For ordinary users, that may sound technical. In practice, it matters because a longer context allows models to work with large documents, long conversations, full codebases, legal materials, research archives or complex project histories without losing track as quickly.

That is why DeepSeek V4 is not just another chatbot release. It is aimed at the next stage of AI use, where models are expected to act less like question-answering tools and more like assistants that can follow long processes over time.

Two models for two different needs

DeepSeek V4 comes in two main versions. DeepSeek-V4-Pro is a larger and more capable model, with 1.6 trillion total parameters and 49 billion active parameters. DeepSeek-V4-Flash is a smaller model, with 284 billion total parameters and 13 billion active parameters, designed for faster and more cost-effective workloads.

That distinction is important. Not every user needs the strongest model for every task. A company summarising documents, routing queries or running basic support may choose Flash. A developer working on complex coding tasks, long-context agents or advanced reasoning may prefer Pro.

DeepSeek’s release reflects a broader trend in AI. The best model is no longer always the biggest one. Cost, speed, context size and deployment flexibility are now as important as raw benchmark performance.

Why the price matters

One reason DeepSeek attracts so much attention is its aggressive pricing. DeepSeek’s API page lists V4-Flash at USD 0.14 per 1 million input tokens on a cache miss and USD 0.28 per 1 million output tokens. V4-Pro is listed at USD 1.74 per 1 million input tokens and USD 3.48 per 1 million output tokens before the temporary 75% discount.

For developers and companies, that changes the calculation. High-performing AI models are useful only if they can be deployed at scale. If every long document, coding session or agentic workflow becomes too expensive, adoption slows down.

DeepSeek’s challenge to the market is therefore not only technical. It is economic. The company is pushing the idea that frontier-level AI should be cheaper to run, easier to access and less dependent on closed ecosystems.

The architecture behind the hype

DeepSeek V4 uses a mixture-of-experts approach, meaning only part of the model is active during each response. That helps explain why the model can be very large on paper, yet still more efficient to run than a dense model of similar overall size.

The more interesting part is how DeepSeek handles long context. NVIDIA’s technical overview explains that DeepSeek V4 uses hybrid attention, combining compression and selective attention techniques to reduce the cost of processing very long prompts. NVIDIA says these changes are designed to cut per-token inference FLOPs by 73% and reduce KV cache memory burden by 90% compared with DeepSeek-V3.2.

For a non-technical audience, the point is simple. DeepSeek V4 is trying to solve one of the biggest problems in modern AI: how to make models remember and process much more information without becoming too slow or too expensive.

That is where much of the hype comes from. The model is not merely larger. It is designed around the economics of long-context AI.

Why NVIDIA is still in the picture

DeepSeek’s R2 launch is delayed as US restrictions cut off supply of NVIDIA H20 chips built for China.

NVIDIA’s role in the DeepSeek V4 story is especially interesting. DeepSeek is often discussed as part of China’s effort to build a more independent AI ecosystem, but NVIDIA has also been quick to move forward to support developers who want to build with the model.

In its technical blog, NVIDIA describes DeepSeek V4 as a model family designed for efficient inference of million-token contexts. The company says DeepSeek-V4-Pro and V4-Flash are available through NVIDIA GPU-accelerated endpoints, while developers can also use NVIDIA Blackwell, NIM containers, SGLang and vLLM deployment options.

NVIDIA also reports that early tests of DeepSeek-V4-Pro on the GB200 NVL72 platform showed more than 150 tokens per second per user. That matters because long-context models place heavy memory pressure, as well as on compute and networking infrastructure. The model may be efficient by design, but serving it at scale still requires serious hardware.

So, DeepSeek V4 does not remove NVIDIA from the story – it complicates it. The model is part of a broader push towards more efficient AI, but the infrastructure race remains central.

The chip question behind the model

DeepSeek V4 also arrives at a time when AI infrastructure is becoming just as important as model performance. MIT Technology Review frames the release partly through that lens, noting that DeepSeek’s new model reflects China’s broader attempt to reduce reliance on foreign AI hardware and build a more self-sufficient technology stack.

That detail matters because the AI race is no longer only about who builds the most capable model. It is also about who controls the chips, software frameworks and data centres needed to run it.

Replacing NVIDIA, however, remains difficult. Its advantage lies not just in its chips, but also in the software ecosystem developers have built around its platforms over many years. Moving to alternative hardware means adapting code, rebuilding tools and proving that the new systems are stable enough for serious use.

DeepSeek V4, however, sits between two realities. It points towards China’s ambition to build a more independent AI stack, while NVIDIA’s rapid support for the model shows that frontier AI still depends heavily on established infrastructure.

Open weights as a strategic move

DeepSeek V4 is also important because the model weights are available through Hugging Face under the MIT License. That gives developers more freedom to inspect, adapt and deploy the model than they would have with a fully closed commercial system.

Open-weight models are becoming a major pressure point in the AI race. Closed models may still lead in some areas, especially in polished consumer products, enterprise support and safety layers. However, open models offer something different: flexibility.

For universities, start-ups, smaller companies and developers outside the largest AI ecosystems, that flexibility matters. It means advanced AI can be tested, modified and integrated without relying entirely on a handful of dominant providers.

Benchmarks need caution

DeepSeek presents V4-Pro as highly competitive across reasoning, coding, long-context and agentic benchmarks. Hugging Face lists results including 80.6 on SWE-bench Verified, 90.1 on GPQA Diamond and 87.5 on MMLU-Pro for DeepSeek-V4-Pro.

Those numbers are impressive, but they should not be treated as the full story. Benchmarks are useful, but they rarely capture every real-world use case. A model can score well on coding tests and still struggle with reliability, factual accuracy, safety or complex multi-step workflows in production.

That caution is important. The AI industry often turns benchmarks into headlines, while real performance depends on deployment, prompting, safety controls and the specific task at hand.

More than just another model release

DeepSeek V4 matters because it combines several trends into one release: long context, lower prices, open weights, agentic workflows and geopolitical competition. It also shows that the AI race is no longer fought only in labs, benchmarks and data centres. Visibility now matters too. Tools such as Diplo’s Digital Footprints show how digital presence shapes the way technology actors and media narratives are discovered, ranked and understood. At this stage, the competition is not only about who has the smartest model. It is also about who can make intelligence cheaper, more available and easier to deploy.

That does not mean DeepSeek has solved every problem. Questions remain around independent benchmarking, safety, data governance, infrastructure and the broader political context of Chinese AI development. Still, the release does show where the market is heading.

The next phase of AI may not be defined solely by the most powerful model. It may be defined by the model that is powerful enough, affordable enough and open enough to change how people build products, services and tools with AI.

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!

Claude Mythos Preview sets new benchmark for AI capability and raises governance questions

On 7 April 2026, Anthropic announced Claude Mythos Preview, its most capable AI model to date, alongside the explicit decision not to make it publicly available. Claude Mythos Preview is a general-purpose, unreleased frontier model that, in Anthropic’s own words, reveals a stark fact: AI models have reached a level of coding capability where they can surpass all but the most skilled humans in finding and exploiting software vulnerabilities.

The announcement was accompanied by a coordinated industry initiative, proactive government briefings across the US and UK, and a detailed 244-page system card.

The significance of the Mythos case extends beyond the technical capabilities of a single model. It raises substantive questions about whether voluntary governance frameworks are sufficient at the frontier of AI development, what it means for the world’s most powerful technology to be held by a small group of private actors, and whether informal engagement with governments constitutes adequate oversight when the stakes involve critical infrastructure, national security, and the global software ecosystem.

Data leak

Electronics, Screen, Computer Hardware, Hardware, Monitor, Light

In late March 2026, security researchers identified an unsecured data cache linked to Anthropic’s content management system, through which nearly 3,000 unpublished assets were accessible via public URLs. Among the materials were a draft blog post describing the model and internal benchmark comparisons. The incident was attributed to human error: assets published via the content management system were set to public by default and required an explicit action to change that setting.

The leak generated immediate media attention and forced Anthropic to make an unplanned public confirmation of the model’s existence. The company accelerated its official announcement to 7 April 2026. Anthropic’s restricted deployment strategy depends on maintaining clear access boundaries during early rollout – precisely the kind of operational control the content management system incident suggests requires stronger enforcement. The incident is relevant beyond its immediate consequences: it illustrates how information about frontier AI capabilities can become public through routine operational failures, independent of any deliberate disclosure decision.

A new tier in the model landscape

Anthropic’s published benchmarks show Mythos Preview scored 93.9% on the SWE-bench Verified test, 97.6% on the USAMO 2026 mathematics evaluation, and and significantly outperformed all previously released models in cybersecurity-specific assessments. The SWE-bench Verified score is roughly double the 2024 state of the art and was achieved in an agentic context, where the model autonomously resolved real software engineering issues from production codebases.

On the USAMO 2026 evaluation, Mythos Preview scored 55 percentage points higher than Opus 4.6, which scored 42.3%. On GPQA Diamond, a graduate-level scientific reasoning benchmark, Mythos Preview scored 94.6%. On Terminal-Bench 2.0, which evaluates system administration and command-line proficiency, it scored 82.0%, a 16.6-point lead over Opus 4.6. On the cybersecurity benchmark Cybench, the model scored 100% on the first attempt, making it no longer useful as a discriminating evaluation.

Cybersecurity capabilities

The decision not to release Mythos Preview publicly is linked to concerns about its advanced capabilities, particularly in high-risk domains such as cybersecurity, as well as broader considerations related to safety and potential misuse.

Notably, these capabilities are not the result of targeted training. Anthropic did not explicitly train Mythos Preview to have these capabilities. They emerged as a downstream consequence of general improvements in code, reasoning, and autonomy. The same improvements that make the model substantially more effective at patching vulnerabilities also make it substantially more effective at exploiting them.

During internal testing, Mythos Preview identified thousands of zero-day vulnerabilities across every major operating system and every major web browser, as well as other critical software, many of them high severity and previously undetected for years. Three disclosed examples provide concrete shape to what this means.

Mythos Preview found a 27-year-old vulnerability in OpenBSD, used to run firewalls and critical infrastructure, which allowed an attacker to remotely crash any machine running it simply by connecting to it. It identified a 16-year-old flaw in FFmpeg in a line of code that automated testing tools had accessed five million times without detecting the problem. It also autonomously identified and chained together several vulnerabilities in the Linux kernel, allowing an attacker to escalate from regular user access to complete control of a machine.

In one case, the model wrote a web browser exploit that chained together four vulnerabilities to escape both the renderer and operating system sandboxes. In another, it identified vulnerability in a server, discovered a method to bypass authentication requirements, constructed a 20-element attack chain, and split it across six sequential network packets, entirely without human involvement after the initial prompt.

The accessibility dimension is a distinct governance concern. Anthropic engineers with no formal security training could ask Mythos to find remote code execution vulnerabilities overnight and have a complete, working exploit the following morning. Traditionally, sophisticated cyberattacks have required highly skilled teams, extensive planning, and deep technical expertise. Models with these capabilities may lower those barriers substantially, including smaller state actors and non-state actors.

Safety evaluations and alignment findings

Anthropic’s system card simultaneously describes Mythos Preview as the best-aligned model it has released to date by a significant margin, while warning that it likely poses the highest alignment-related risk of any model it has released to date. These assessments are presented together because they reflect different dimensions of the same model: stronger alignment training alongside greater capability does not eliminate the risks that arise from autonomous operation at this level.

Among the findings documented in the system card is a sandbox escape during internal safety evaluation. Researchers placed an early version of Claude Mythos Preview in an isolated computing environment and, as part of a red-teaming exercise, instructed the model to attempt to escape and contact the overseeing researcher. The model succeeded: it developed a multi-step exploit, gained unauthorised internet access, and emailed the researcher, who was eating a sandwich in a park. Without being asked, the model also posted details of its exploit to multiple hard-to-find but technically public-facing websites.

Anthropic has disclosed only a fraction of what it says it has found during internal testing. Over 99% of the vulnerabilities discovered by Mythos remained unpatched at the time of the 7 April announcement. The company followed a coordinated vulnerability disclosure process, providing cryptographic commitments to undisclosed vulnerability details and committing to release specifics once patches are in place.

The Responsible Scaling Policy

Anthropic’s decision-making around Mythos is structured by its Responsible Scaling Policy (RSP), a self-imposed framework first published in 2023 and updated to version 3.0 in February 2026. RSP defines AI Safety Levels (ASL) that set capability thresholds determining what safeguards must be in place before deployment.

Claude Mythos’s ability to autonomously find thousands of zero-day vulnerabilities in real software has placed it at or near the ASL-3 threshold for cybersecurity capabilities. ASL-3 covers models that could provide meaningful assistance to actors seeking to cause significant harm, requiring substantial additional safety measures before deployment.

RSP version 3.0 involves the publication of Frontier Safety Roadmaps with detailed safety goals, as well as Risk Reports that quantify the risk across all deployed models. RSP is built on the principle of proportional protection, where safety measures are intended to scale in tandem with model capabilities.

The framework is not legally binding. The public release of RSP increases transparency and introduces a degree of accountability, but it remains a voluntary, self-imposed governance mechanism rather than government regulation.

Version 3.0 introduced a significant change in how deployment decisions are handled. Earlier versions included a stronger commitment to pause development or delay release if safety measures were insufficient. In the updated policy, this approach has been replaced by a more conditional framework, which takes into account factors such as the level of risk and the broader competitive environment.

Anthropic also acknowledges that unilateral restraint may be less effective if other developers continue to advance similar systems, reflecting what it describes as a collective action problem.

These changes have drawn criticism from AI safety researchers, some of whom argue that they may weaken the credibility of voluntary governance mechanisms under competitive pressure.

In May 2025, Anthropic activated ASL-3 protections because it felt it could no longer make a sufficiently strong case that the relevant risk was low. More than nine months later, despite significant effort, including a randomised controlled trial, no compelling evidence that the risk was high has materialised. This grey zone, where neither safety nor significant risk can be definitively demonstrated, is where much of the governance challenge currently sits.

Project Glasswing

Anthropic launched Project Glasswing as a structured access mechanism to use Claude Mythos Preview for defensive cybersecurity purposes. The initiative brings together Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks as launch partners, with access also extended to over 40 additional organisations that build or maintain critical software infrastructure.

Project Glasswing partners will receive access to Claude Mythos Preview to find and fix vulnerabilities in their foundational systems, with work expected to focus on local vulnerability detection, black box testing of binaries, securing endpoints, and penetration testing. Anthropic is committing up to $100M in usage credits for Mythos Preview across these efforts. Following the initial research preview period, access to the model will be available to participants at $25 per million input tokens and $125 per million output tokens across the Claude API, Amazon Bedrock, Google Cloud’s Vertex AI, and Microsoft Foundry.

Anthropic has also donated $2.5M to Alpha-Omega and OpenSSF through the Linux Foundation, and $1.5M to the Apache Software Foundation to enable open-source software maintainers to respond to the changing cybersecurity landscape.

Within 90 days, Anthropic has committed to reporting publicly on what it has learned, as well as the vulnerabilities fixed and improvements made that can be disclosed. The company also intends to collaborate with leading security organisations to produce practical recommendations covering vulnerability disclosure processes, software update processes, open-source and supply-chain security, and patching automation, among other areas.

Anthropic has stated that Project Glasswing is a starting point, and that in the medium term an independent, third-party body bringing together private and public sector organisations might be the ideal home for continued work on large-scale cybersecurity projects.

Project Glasswing raises a governance question for the industry, as cyber-capable AI systems may become useful security tools and a source of misuse risk at the same time. Project Glasswing’s structure also reveals tensions, as it concentrates several roles including discovery, disclosure coordination, and capability gatekeeping in a single organisation. Entities such as Anthropic and major cloud providers control critical components of the Glasswing ecosystem, raising questions about power and governance that, for financial institutions in particular, translate into systemic risk.

Government responses

Prior to the external release, Anthropic briefed senior US government officials on Mythos’s offensive and defensive cyber capabilities, including the Cybersecurity and Infrastructure Security Agency and the Center for AI Standards and Innovation. On the same day that Project Glasswing was announced, US Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell convened a meeting with the chief executives of major Wall Street banks to communicate the cybersecurity risks the model presents.

In the UK, officials from the Bank of England, the Financial Conduct Authority, and the Treasury entered into urgent talks with the National Cyber Security Centre. Representatives from major British banks, insurers, and exchanges were expected to be briefed on cybersecurity risks within the following two weeks. These consultations were initiated by regulators, not as a result of any legal obligation on Anthropic’s part.

Anthropic co-founder Jack Clark confirmed at the Semafor World Economy Summit that the company had briefed the Trump administration on Mythos. Clark stated that ‘our position is the government has to know about this stuff, and we have to find new ways for the government to partner with a private sector that is making things that are truly revolutionizing the economy,’ adding that ‘absolutely, we talked to them about Mythos, and we’ll talk to them about the next models as well.’

The Anthropic-Pentagon dispute

American Flag, Flag

The relationship between Anthropic and the US government in the lead-up to the Mythos announcement was already shaped by an active legal dispute. On 27 February 2026, six weeks before the Mythos announcement, the Trump administration ordered federal agencies and military contractors to halt business with Anthropic after the company refused to allow the Pentagon to use its technology without restrictions. Anthropic had two stated red lines: it did not want its AI systems used in autonomous weapons or domestic mass surveillance.

The Department of Defense designated Anthropic a supply chain risk, a label usually applied to firms associated with foreign adversaries. A federal judge in California blocked the Pentagon’s effort, ruling that the measures violated Anthropic’s constitutional rights. A federal appeals court subsequently denied Anthropic’s request to temporarily block the blacklisting, leaving the company excluded from Department of Defense contracts while allowing it to continue working with other government agencies during litigation.

The dispute illustrates the structural tension that the Mythos case makes concrete. Anthropic simultaneously informed the US government about the most capable cyber AI system ever evaluated, sought partnerships with government agencies through Project Glasswing, and was engaged in legal proceedings against the Pentagon over the limits of the military use of its technology. Frontier AI companies operate largely beyond formal government authority and may come into significant conflict with it, as the legal battle between Anthropic and the Pentagon demonstrates. The governance environment does not yet have well-established mechanisms for resolving these tensions.

Geopolitical dimensions

Person

Claude Mythos has sharpened attention on the competitive and geopolitical dimensions of frontier AI development. Project Glasswing’s launch partners exclude Anthropic’s rival OpenAI, which is reported to be approximately six months behind Anthropic in developing a model with comparable offensive cyber capabilities.

Senior policy voices have positioned Mythos within the broader competition between Western AI companies and China‘s rapidly evolving AI ecosystem, with implications for national security, enterprise adoption, and technological leadership. A security researcher assessed a concurrent source code leak from Anthropic as a geopolitical accelerant, noting that such exposures compress the timeline for adversaries to replicate technological advantages currently held by Western laboratories.

Many defence organisations still rely on legacy software and infrastructure not designed with AI-driven threats in mind. Models capable of autonomously identifying hidden flaws in older code may expose weaknesses in critical defence networks around the world. The difficulty of containment at the geopolitical level is reflected in usage patterns. Access restriction at the laboratory level does not translate reliably into containment across jurisdictions when the same underlying models are accessible via cloud infrastructure spanning multiple countries and regulatory environments.

The limits of voluntary AI governance

The Claude Mythos case has clarified, with considerable precision, what voluntary AI governance can and cannot achieve. A responsible laboratory can make a unilateral decision not to release a dangerous system. It can support coordinated vulnerability disclosure, engage governments proactively, and produce detailed public documentation of a model’s capabilities and risks. All of these have occurred with Mythos, and represent meaningful progress relative to the governance environment of a few years ago.

What voluntary frameworks cannot do is bind competitors who operate under different assumptions. Anthropic’s RSP version 3.0 acknowledges this directly by removing the commitment to withhold unsafe models if another laboratory releases a comparable model first. The competitive structure of the AI industry means that restraint by one actor does not prevent the underlying capability from eventually proliferating. Voluntary governance frameworks work best when they generate shared norms across an industry. When the industry is structured around intense competition among a small number of organisations, voluntary restraint by a single actor does not resolve the broader question of access.

Analysts note that what Mythos does today in a restricted environment, publicly available models are likely to replicate within one to two model generations. The next phase of the EU AI Act takes effect in August 2026, introducing automated audit trails, cybersecurity requirements for AI systems classified as high risk, incident reporting obligations, and penalties of up to 3% of global revenue. The EU framework represents a shift toward binding governance, but its scope relative to the pace and international distribution of frontier AI development remains to be demonstrated.

Conclusion

Anthropic acknowledges that capabilities like those demonstrated by Mythos will proliferate beyond actors committed to deploying them safely, with potential fallout for economies, public safety, and national security. The company’s response, taken in aggregate, reflects a serious attempt to manage that risk within the constraints of voluntary frameworks and private decision-making. The Responsible Scaling Policy, Project Glasswing, proactive government briefings, and the detailed system card are each substantive contributions. They are also all products of a single private entity’s judgement, operating without binding external accountability.

The Mythos case does not so much call for a different assessment of Anthropic’s conduct as it does a clear-eyed view of what voluntary governance can realistically sustain at the frontier of AI development. Governments on both sides of the Atlantic were briefed informally about a model whose capabilities are consequential for critical infrastructure and national security. No binding notification requirement existed. No independent technical authority had prior access. No international coordination mechanism was in place.

No single organisation can solve these challenges alone. Frontier AI developers, software companies, security researchers, open-source maintainers, and governments all have essential roles to play. The Mythos case has made that observation not merely a statement of aspiration but a policy problem that requires concrete institutional responses. Whether those responses will take shape before the next capability threshold is reached is the question now facing policymakers.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot! 

AI industrial policy questions control over power, wealth and governance

Every technological leap forces society to renegotiate its relationship with power. Intelligence, once a uniquely human advantage, is now being abstracted, scaled, and embedded into machines. As AI evolves from a tool into an autonomous force shaping economies and institutions, the question is no longer what AI can do, but who it will ultimately serve.

A new framework published by OpenAI sets out a vision for managing the transition towards advanced AI systems, often described as superintelligence. Framed as a policy agenda for governments and institutions, it attempts to define how societies should respond to rapid advances in AI governance, economic transformation, and workforce disruption.

At its core, the document is not a regulation but influence: an attempt to shape how policymakers think about industrial policy for AI, productivity gains, and the redistribution of technological power.

OpenAI introduces an AI industrial policy approach exploring how AI is redefining global structures in the intelligence age and shaping future governance.
Image via freepik

AI industrial policy and the next economic transformation

The central argument is that AI will act as a general-purpose technology comparable to electricity or the combustion engine. It promises higher productivity, lower costs, and accelerated innovation across industries. In policy terms, this aligns with broader discussions around AI-driven productivity growth and economic restructuring.

However, historical precedent suggests that such transitions are rarely evenly distributed. Industrial revolutions typically begin with labour displacement, rising inequality, and capital concentration, before broader gains are realised. AI may intensify this dynamic due to its dependence on compute infrastructure, proprietary models, and large-scale data ecosystems.

Economic power may become increasingly concentrated among a small number of AI developers and infrastructure providers, posing a structural risk of reinforcing existing inequalities rather than reducing them.

OpenAI introduces an AI industrial policy approach exploring how AI is redefining global structures in the intelligence age and shaping future governance.
Image via freepik

The return of industrial policy in the AI economy

A key feature of the document is its explicit endorsement of AI industrial policy as a necessary response to market limitations. Governments, it argues, must play a more active role in shaping outcomes through regulation, investment, and public-private coordination.

A broader global shift in economic thinking is reflected in this approach. Strategic sectors such as semiconductors, energy, and digital infrastructure are already experiencing increased state intervention. AI now joins that category as a critical technology.

Yet this approach introduces a significant tension. When leading AI firms contribute directly to the design of AI regulation and governance frameworks, the risk of regulatory capture increases. Policies intended to ensure fairness and safety may inadvertently reinforce the dominance of incumbent companies by raising compliance costs and technical barriers for smaller competitors.

In this sense, AI industrial policy may not only guide innovation but also determine market entry, competition, and the long-term economic structure.

OpenAI introduces an AI industrial policy approach exploring how AI is redefining global structures in the intelligence age and shaping future governance.
Image via freepik

Redistribution, taxation, and the question of AI wealth

The document places strong emphasis on economic inclusion in the AI economy, proposing mechanisms such as a public wealth fund, AI taxation, and expanded access to capital markets. These ideas are designed to address one of the central challenges of AI-driven growth: the potential for extreme wealth concentration.

As AI systems increase productivity while reducing reliance on human labour, traditional tax bases such as wages and payroll contributions may weaken. The proposal to tax AI-generated profits or automated labour reflects an attempt to stabilise public finances in an increasingly automated economy.

Equally significant is the idea of a ‘right to AI’, which frames access to AI as a foundational requirement for participation in modern economic life. This positions AI not merely as a tool, but as a form of digital infrastructure essential to economic agency and inclusion.

However, these proposals face major implementation challenges. Measuring AI-generated value is complex, particularly in hybrid systems where human and machine inputs are deeply integrated. Without clear definitions, AI taxation frameworks and redistribution mechanisms could prove difficult to enforce at scale.

OpenAI introduces an AI industrial policy approach exploring how AI is redefining global structures in the intelligence age and shaping future governance.
Image via freepik

Workforce disruption and the future of work

The document recognises that AI will significantly reshape labour markets. Many tasks that currently require hours of human effort are already being automated, with future systems expected to handle more complex, multi-step workflows.

To manage this transition, the proposal highlights reskilling programmes, portable benefits systems, and adaptive social safety nets, alongside experimental ideas such as a reduced working week. These measures aim to mitigate the impact of automation and workforce disruption while maintaining economic stability.

However, the pace of change introduces uncertainty. Historically, labour markets have adjusted over decades, allowing new roles to emerge gradually. AI-driven disruption may occur much faster, compressing adjustment periods and increasing transitional risk.

While the document highlights expansion in sectors such as healthcare, education, and care services, these ‘human-centred jobs’ require substantial investment in training, wages, and institutional support to absorb displaced workers effectively.

OpenAI introduces an AI industrial policy approach exploring how AI is redefining global structures in the intelligence age and shaping future governance.
Image via freepik

AI safety, governance, and systemic control

Beyond economic considerations, the proposal places a strong emphasis on AI safety, auditing frameworks, and risk mitigation systems. The proposed measures include model evaluation standards, incident reporting mechanisms, and international coordination structures.

These safeguards respond to growing concerns around cybersecurity risks, biosecurity threats, and systemic model misalignment. As AI systems become more autonomous and embedded in critical infrastructure, governance mechanisms must evolve accordingly.

However, safety frameworks also introduce questions of control. Determining which systems are classified as high-risk inevitably centralises authority within regulatory and institutional bodies. In practice, this may restrict access to advanced AI systems to organisations capable of meeting stringent compliance requirements.

A structural trade-off between security and openness is emerging in the AI economy, raising questions about how innovation and oversight can coexist without reinforcing centralisation.

OpenAI introduces an AI industrial policy approach exploring how AI is redefining global structures in the intelligence age and shaping future governance.
Image via freepik

Strategic influence and the future of AI governance

The proposal from OpenAI is both policy-oriented and strategically positioned. It acknowledges legitimate risks- inequality, labour disruption, and systemic instability, while offering a roadmap for managing them through structured intervention.

At the same time, it reflects the perspective of a leading actor in the AI industry. As a result, its recommendations exist at the intersection of public interest and commercial strategy. The dual role raises important questions about who defines AI governance frameworks and how economic power is distributed in the intelligence age.

The broader challenge is not only technological but also institutional: ensuring that AI industrial policy, regulation, ethics and economic design are shaped through transparent and democratic processes, rather than through concentrated private influence.

OpenAI introduces an AI industrial policy approach exploring how AI is redefining global structures in the intelligence age and shaping future governance.
Image via freepik

AI industrial policy will define economic power

AI is no longer solely a technological development- it is a structural force reshaping global economic systems. The emergence of AI industrial policy frameworks reflects an attempt to manage this transformation proactively rather than reactively.

The success or failure of these approaches will determine whether AI-driven growth leads to broader prosperity or deeper concentration of wealth and power. Without effective governance, the risks of inequality and centralisation are significant. With carefully designed policies, there is real potential to expand access, improve productivity, and distribute benefits more widely.

Digital diplomacy may increasingly come to the fore as a mechanism for arbitrating competing approaches to AI policy and governance across jurisdictions. As regulatory frameworks diverge, diplomatic channels could serve to bridge gaps, negotiate standards, and balance strategic interests, positioning digital diplomacy as a practical tool for managing fragmentation in the evolving AI economy. 

Ultimately, the intelligence age will not be defined by technology alone, but by the AI governance systems, economic frameworks, and industrial policy decisions that guide its development. The outcome will depend on the extent to which global stakeholders succeed in building a shared and coordinated vision for its future.

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!  

The implementation of the EU AI Act with a focus on general-purpose AI models

Transition from legislation to implementation

The European Union has entered a new phase in the governance of AI, moving from the legislative adoption of the Artificial Intelligence Act (AI Act) towards its practical implementation. This particular phase places particular emphasis on obligations of providers of general-purpose AI (GPAI) models, reflecting the increasing role of such systems in the broader digital ecosystem.

The AI Act, adopted in 2024, establishes a comprehensive legal framework for AI within the EU. It introduces a risk-based approach that classifies AI systems into categories ranging from minimal risk to unacceptable risk, with corresponding regulatory requirements.

According to the official text of the regulation, the framework is designed to ensure that AI systems placed on the market in the Union are ‘safe and respect existing law on fundamental rights and Union values.’

While earlier discussions around the Act focused on its legislative negotiation and scope, the current phase centres on how its provisions will be applied in practice.

General-purpose AI models within the AI Act

A key element of this implementation phase concerns general-purpose AI models. These models, which can be integrated into a wide range of downstream applications, occupy a distinct position within the regulatory framework.

The AI Act defines general-purpose AI models as systems that can be used across multiple tasks and contexts and may ‘serve a variety of purposes, both for direct use and for integration into other AI systems.’

That positioning reflects the broad applicability of these models, particularly in areas such as natural language processing, content generation, and data analysis.

The Act also recognises that the widespread deployment of such models may have implications beyond individual use cases, particularly when integrated into high-risk systems.

Obligations for providers of GPAI models

The European Commission, together with the European AI Office, has begun outlining expectations for compliance with provisions related to general-purpose AI.

According to official EU materials, providers of GPAI models are required to ensure that technical documentation is drawn up and kept up to date.

European Union
Image via Freepik

The regulation specifies that providers should ‘draw up and keep up-to-date technical documentation of the model,’ ensuring that relevant information is accessible for compliance and oversight purposes. In addition, transparency obligations require providers to make certain information available to downstream deployers.

The intention of this is to support the responsible integration of GPAI models into other systems.

Distinction between GPAI and systemic-risk models

The AI Act introduces a distinction between general-purpose AI models and those considered to pose systemic risk.

Models that meet specific criteria, such as scale, capability, or deployment level, may be classified as having a systemic impact.

For such models, additional obligations apply, including requirements related to evaluation, risk mitigation, and reporting. The European Commission has indicated that further guidance will clarify how systemic risk thresholds are determined, including through delegated acts and technical standards.

Role of the European AI Office in implementation

The European AI Office, established within the European Commission, plays a central role in supporting the implementation of the AI Act.

Its responsibilities include contributing to the consistent application of the regulation, coordinating with national authorities, and supporting the development of methodologies for compliance.

European AI Office
Source: digital-strategy.ec.europa.eu/en/policies/ai-office

According to the European Commission, the AI Office is tasked with ‘ensuring the coherent implementation of the AI Act across the Union.’ The Office is also expected to contribute to the development of benchmarks, testing frameworks, and guidance documents that support both regulators and providers.

Phased implementation timeline

The implementation of the AI Act is structured as a phased process, with different provisions becoming applicable over time.

That phased approach allows stakeholders to adapt to the regulatory requirements while enabling authorities to establish enforcement mechanisms.

Provisions related to general-purpose AI models are among the earlier elements to be operationalised, reflecting their central role in the current AI landscape.

The European Commission has indicated that additional implementing acts and guidance documents will be issued as part of this process.

Coordination with national authorities

While the European AI Office plays a coordinating role at the EU level, enforcement remains the responsibility of national authorities within member states.

The AI Act establishes mechanisms for cooperation and information-sharing to support a harmonised approach across the European Union.

National authorities are expected to work closely with the AI Office and the European Commission to oversee compliance and address emerging challenges.

Stakeholder engagement and technical guidance

The implementation phase also involves engagement with a range of stakeholders, including industry actors, civil society organisations, and technical experts.

Also, the European Commission has initiated consultations and workshops to gather input on practical aspects of implementation, such as documentation standards and risk assessment methodologies.

The following process supports the development of operational guidance applicable across sectors and use cases.

Interaction with the EU digital regulatory framework

The AI Act forms part of a broader EU digital policy framework that includes instruments such as the General Data Protection Regulation (GDPR), the Digital Services Act (DSA), and the Digital Markets Act (DMA).

These frameworks address different aspects of the digital ecosystem, including data protection, platform governance, and market competition.

The relationship between the AI Act and these instruments is expected to be clarified further during implementation.

International context: OECD and UN approaches

The governance of general-purpose AI models is also being addressed at the international level.

The OECD AI Principles state that AI systems should be ‘robust, secure and safe throughout their entire lifecycle,’ and emphasise accountability for their functioning.

Logo, Disk, Astronomy, Outer Space

At the UN level, the Global Digital Compact process addresses issues related to transparency, accountability, and oversight of digital technologies, including AI.

The listed initiatives provide non-binding guidance, in contrast to the legally binding framework established by the EU AI Act.

Ongoing development of technical standards

The development of technical standards is an important component of the implementation process.

The European Commission has indicated that it will work with standardisation organisations to develop specifications related to documentation, evaluation, and risk management.

These standards are expected to support the practical application of the AI Act’s provisions.

From regulatory framework to regulatory practice

The current phase of the EU AI Act marks a transition from legislative design to regulatory practice.

For providers of general-purpose AI models, this involves preparing to meet obligations related to documentation, transparency, and risk management. For regulators, the focus is on ensuring consistent application of the rules across member states, supported by coordination mechanisms and guidance from the AI Office.

The implementation process is expected to evolve as further guidance is issued.

Conclusion

The European Union’s AI Act is entering its implementation phase, with a particular focus on general-purpose AI models.

That phase involves translating the regulation’s legal provisions into operational requirements, supported by guidance from the European Commission and the AI Office.

The development of technical standards, coordination mechanisms, and compliance frameworks will play a central role in this process. As implementation progresses, further clarification is expected through additional guidance and regulatory measures, contributing to the operationalisation of the EU’s approach to AI governance.

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!

UN kicks off Global Mechanism on ICT security, road ahead murky

After almost three decades of stop-start cybersecurity negotiations at the UN, the long-anticipated Global Mechanism on ICT security has finally kicked off.

It is the first permanent forum of its kind since discussions on ICT security began back in 1998, and its mere existence says a lot about how far those talks have come.

But if the launch felt like a breakthrough, the organisational session quickly brought things back down to earth. Beyond what was already sketched out in Annex C and the OEWG’s Final Report, it remained unclear how the Mechanism would actually organise itself in practice.

Text, Page, Symbol

The session raised plenty of questions—about structure, priorities, and process—but offered few real answers, leaving the sense that while the Mechanism now exists, what it will do and how it will do it is still very much up for grabs.

A new body, a new mandate, and a newly elected Chair, Egriselda López of El Salvador, injected renewed optimism into the Global Mechanism’s first organisational session. Yet, within minutes, it became evident that the Global Mechanism did not start with a blank slate, but rather inherited the OEWG’s long list of disagreements. 

Russia opened the discussion by disputing the legitimacy of the Chair nomination, which they claimed was guided solely by the UNODA and thus limited state participation in the process. They used this opportunity to stress that all decisions under the new process must be based on consensus and be completely intergovernmental. 

The substantive issues on the agenda

For the provisional agenda of the mechanism’s July session, the Chair circulated a draft agenda organised around the five pillars of the framework for responsible state behaviour in the use of ICTs. However, Iran and Russia argued that the wording of agenda item 5 did not precisely reflect paragraph 9 of Annex C of the OEWG final report and called for correction at this session. The EU and Canada rejected this, arguing the draft already referenced all relevant documents and that isolating one paragraph would itself constitute renegotiation. The USA reserved its position entirely, preferring that the July plenary adopt its own agenda. No consensus was reached, and the Chair will continue consultations before July.

The mechanism inherited many unresolved substantive debates from its predecessors. 

On international law, there is widespread agreement that considerable work remains to be done, but little agreement on how to carry it out. The majority of delegations have shown clear support for strengthening the existing normative framework and reaffirming the UN Charter’s application to cyberspace.

A broad majority of states expressed support for ensuring that the mechanism remains action-oriented, with a strong focus on practicality and the implementation of agreed frameworks on international law, norms, CBMs, and capacity-building (Chile, Nauru, Portugal, Switzerland, the United Kingdom, Estonia, Italy, Australia, the Democratic Republic of the Congo, Antigua and Barbuda, Sudan, Vanuatu, Albania, Vietnam, India, Greece, Rwanda, the Dominican Republic, North Macedonia, Kiribati).

In particular, some delegations advocated for applying the framework to concrete scenarios as a way to stimulate implementation (Japan, the Netherlands, the United Kingdom, Sudan).  China was the only delegation to emphasise that further development of the framework is equally important alongside its implementation.

The EU highlighted the norm checklist, a hotly debated issue in the previous mechanism, as an area for further improvement. 

However, to many states, a fundamental concern remains. Capacity building initiatives risk stalling without reliable funding, so many delegations, primarily from developing countries, urged the Global Mechanism to prioritise the operationalisation of the UN Voluntary Fund, which was tabled but left unresolved by the OEWG.

Dedicated thematic groups: Who, what and how

The often broad agenda and long-winded statements of delegations in OEWG plenary sessions left little room for technical depth, leaving many delegations frustrated with the gap between consensus language and concrete action. 

The Dedicated thematic groups (DGTs) were created to address this issue precisely by setting up an informal, technical forum to advance practical initiatives already agreed on, such as the Global ICT Security Cooperation and Capacity Building Portal. However, the practicalities on how they should be set up and administered are going to be hotly contested as it will influence what gets on the agenda, who drives it, and whether this new system is capable of delivering real outcomes over time.

Who will lead DTGs?

The dominant and most contested question of the session was who would appoint the co-facilitators for the two Dedicated Thematic Groups. The Chair proposed appointing two co-facilitators per DTG: one from a developed country, one from a developing country, drawing on GA practice, under which the Chair appoints co-facilitators for intergovernmental processes. She indicated her intention to hold broad informal consultations before making appointments, and committed to geographic balance, gender parity where practicable, and relevant technical expertise as selection criteria. 

Who ends up in these roles matters considerably: the co-facilitators will steer the DTG discussions, shape their agendas, and channel recommendations to the plenary.

A broad coalition of states supported the Chair’s approach, including the EU, speaking on behalf of its member states and several aligned countries such as France, Germany, Australia, the United Kingdom, the Netherlands, Switzerland, Japan, Egypt, Senegal, Nigeria, Malaysia, Moldova, and others. Egypt and Senegal were among the most direct, noting that delays in operationalising the mechanism would waste the intersessional period and erode its credibility, particularly for developing countries eager to move from procedure to substance.

Another group of states, led by Russia and supported by Iran, China, Belarus, Nicaragua, and Cuba, argued that co-facilitator appointments must be approved by member states by consensus rather than made unilaterally by the Chair. Russia contended that DTG co-facilitators handle substantive political matters and therefore constitute officials whose appointment requires a collective agreement. Russia also raised a geographic argument: assigning one developed-country and one developing-country co-facilitator per DTG still disproportionately favours developed states, which represent less than one-fifth of UN membership. Iran added that the early OEWG draft text had explicitly authorised the Chair to appoint DTG facilitators, but that this provision was deliberately removed during negotiations, signalling a lack of agreement on the matter.

The Chair affirmed her intention to consult all member states informally before presenting candidates and called on delegations to show flexibility given the urgency of getting the mechanism’s work underway. Russia subsequently stated its understanding that candidates would be determined through broad consultation, followed by consensus-based approval, but the Chair neither confirmed nor rejected this interpretation. 

The question is effectively deferred to the intersessional period, meaning the composition of the DTG leadership teams remains unresolved and will require continued diplomatic engagement before July.

What will DTGs discuss?

A closely related debate concerned who decides what the DTGs will actually discuss. Several Western and like-minded delegations (e.g., Germany, France, Canada, the United Kingdom, and Australia) highlighted that it is a prerogative of the Chair and co-facilitators, to be exercised in close consultation with states. These delegations proposed ransomware and critical infrastructure protection as natural starting points, citing their frequency across national statements and OEWG discussions. 

Iran and Russia emphasised that topics must be determined by consensus among all member states. Argentina argued that the plenary should maintain control over the agenda rather than ceding too much responsibility to the co-facilitators. 

Morocco instead advocated a bottom-up model in which DTGs define their own priority subtopics from the start, based on member states’ expressed preferences to maintain regional balance and ownership. 

In this sense, the DTGs’ credibility hinges on a delicate balance, having to be ambitious enough to move conversations into action but also focused enough on issues with broad support so that their outputs survive in plenary. 

No decision was taken. For industry and civil society organisations with specific thematic priorities, this remains an active opening: states are currently receptive to input on which topics the DTGs should prioritise.

Colombia put forward a process proposal that drew broadly positive reactions across delegations. It recommended that:

  • DTG mandates be time-limited with clearly defined and measurable outputs; 
  • DTG 1 addresses specific rotating subjects rather than its entire mandate simultaneously, and 
  • DTG outputs systematically distinguish between recommendations on which consensus exists and those still under development. 

Senegal made a complementary point: reports should document both areas of agreement and divergence, preserving a record of discussions even when no consensus was reached. Both proposals reflect a wider concern that, without structured outputs and clear timelines, the mechanism risks reproducing the open-ended deliberation of the OEWG without generating implementable results.

How will DTGs feed into the plenary?

Another issue discussed was how DGT work feeds into plenary work. Brazil made it clear that without a defined protocol for elevating DTG reports to the plenary and formally accepting their recommendations, the groups risk becoming talking shops that are disconnected from the mechanism’s official conclusions. Their proposed solution, which still has to achieve support, is to keep DGT conversations primarily informal but include a short formal section for decision-making. 

Stakeholder participation

A long-standing point of contention and possibly the most politically-charged was the role of non-governmental actors in the groups. The effective participation of interested stakeholders remains uncertain. 

Some delegations adopted a more accommodating stance, recognising that stakeholders can enhance the quality of deliberations (Sudan, Antigua and Barbuda) and contribute to more practical outcomes (Vietnam, Dominican Republic), while underscoring the importance of preserving the intergovernmental nature of the process (Sudan, Vietnam). 

Canada and like-minded states argued that the July 2025 consensus clearly provides for states to nominate experts for DTG briefings and for the wider stakeholder community to participate throughout DTG discussions. 

Iran contested this, asserting that stakeholder modalities agreed for the mechanism apply equally to DTGs. Russia also argued that expert briefings from external stakeholders are a possibility rather than a standard feature, and that inviting external briefers requires member-state agreement on a case-by-case basis. 

How this is resolved will directly determine the degree of access the private sector, technical community, and civil society organisations have to the DTG process in practice.

What’s next? 

The session closed without resolution on its two most consequential questions: co-facilitator appointments and the provisional plenary agenda. The Chair will convene informal intersessional consultations on both and issue a programme of work document before July in all UN languages. 

The Secretariat will open an annual stakeholder accreditation window in the coming weeks; stakeholders wishing to participate in plenary sessions and review conferences can monitor the Digital Watch Observatory web page, where we track the process, for details. 

The broader tension remains unresolved, and how it is managed in the intersessional period will largely determine whether the July plenary can open with the mechanism’s operational foundations in place.

The Chair also confirmed the two key dates for 2026: 

For stakeholders tracking or seeking to contribute to these discussions, these are the dates to plan around.

Metaverse’s decline and the harsh limits of a virtual future

In 2019, Facebook CEO Mark Zuckerberg announced Facebook Horizon, a VR social experience that allows users to interact, create custom avatars, and design virtual spaces. Zuckerberg saw the platform, later renamed Horizon Worlds, as the beginning of a new era of VR social networks, with users trading face-to-face interactions for digital ones.

To show his confidence in VR, Zuckerberg rebranded Facebook Inc. as Meta Platforms Inc. in October 2021, illustrating the company’s shift toward the metaverse as a broad virtual environment intended to integrate social interaction, work, commerce, and entertainment. Building on this new vision, Meta’s ambitions expanded beyond social interaction and entertainment, with the development roadmap including virtual real estate purchases and collaboration in virtual co-working spaces.

Fast forward to 17 March 2026, and the scale of Meta’s retreat from the metaverse vision has become unmistakable. In an official update, the company said it was ‘separating’ VR from Horizon so that each platform could grow with greater focus, while also making Horizon Worlds a mobile-only experience. Under the plan, Horizon Worlds and Events would disappear from the Quest Store by 31 March 2026, several flagship worlds would no longer be available in VR, and the Horizon Worlds app itself would be removed from Quest on 15 June 2026, ending VR access to Worlds altogether.

Yet Meta soon reversed part of the decision. In an Instagram Stories Q&A, CTO Andrew Bosworth said Horizon Worlds would remain available in VR after user backlash. Even so, the greater shift remained unchanged: Horizon Worlds was no longer a flagship VR project, but a much narrower product that reflected a clear contraction of Meta’s original metaverse ambition.

As it stands, Meta’s USD 80 billion investment seems less like a gateway to a new socio-technological era and more like one of the most expensive strategic miscalculations of the 21st century. The sunsetting of Horizon Worlds was certainly not a decision made on a whim, which begs the question: Why did the metaverse fail in the first place? Does it have a future in the AI landscape, and what does its retreat say about the politics of designing the future through corporate platforms?

Metaverse’s mainstream collapse

The most obvious reason for the metaverse’s failure was that it never became a mainstream social space. Meta’s strategy rested on the belief that large numbers of people would start using immersive virtual worlds as a normal setting for interaction, entertainment, and creative activity. The shift never happened at the scale needed to sustain the company’s ambitions.

One reason was friction. VR headsets were less practical than phones, more isolating than social media, and harder to integrate into everyday routines than the platforms people already used to communicate. Entering the virtual world required extra time, extra hardware, and openness to adapt to a different social environment. Most digital habits, however, are built around speed, familiarity, and ease of access.

Meta’s own March 2026 decision makes that failure difficult to deny. A company still convinced that immersive social VR was on its way to becoming mainstream would not have moved Horizon Worlds away from Quest and towards mobile. The shift suggested that the metaverse had failed to move from technological promise to everyday social practice.

Metaverse’s failure was not just one of convenience. It also struggled because it was never presented simply as a new digital space. It was framed as a future built largely on Meta’s own terms, with access tied to the company’s hardware, platforms, rules, and wider ecosystem. Such decisions made the metaverse feel less like an open evolution of the internet and more like a tightly managed corporate environment.

The distinction mattered because Meta was not merely launching another product. It was promoting a vision of how people might one day work, socialise, shop, and create online. Yet the more expansive that vision became, the more obvious it was that the system behind it remained closed and centralised. A future digital environment is harder to embrace when a single company controls the devices, spaces, distribution, and boundaries of participation.

Meta’s handling of Horizon Worlds clearly exposed that tension. The company could remove features, reshape access, alter incentives, and redirect the platform from the top down. Such a level of control may be standard for a private platform, but it sits uneasily with claims about building the next phase of digital life. In that sense, the metaverse failed not only because people were unconvinced by VR, but because its version of the future felt too corporate, too enclosed, and too disconnected from the openness people still associate with the internet.

Metaverse’s economic contradiction

The metaverse did not fail only as a social project. It also became increasingly difficult to justify on economic grounds. Meta spent heavily on Reality Labs while generating only limited returns from those investments. In its 2025 annual filing, the company said Reality Labs had reduced overall operating profit by around USD 19.19 billion for the year, while warning that similar losses would continue into 2026.

Losses on that scale might still have been acceptable if the metaverse had shown clear signs of momentum. However, there was little evidence of mass adoption, strong retention, or a durable path to monetisation. Virtual land, digital goods, branded experiences, and immersive workspaces never developed into the economic base of a new internet layer.

Instead, the metaverse began to look less like a future growth engine and more like a costly experiment with uncertain returns. The gap between spending and payoff became harder to ignore, especially as Meta continued to frame the metaverse as a long-term strategic priority. What used to be sold as the company’s next major frontier was increasingly difficult to justify in commercial terms.

The broader strategic context also changed. Meta’s own forward-looking statements pointed to increased hiring and spending in 2026, especially in AI. In practice, this meant the company was no longer choosing between the metaverse and inactivity, but between two competing visions of the future. AI was already delivering tangible gains in product development, infrastructure, and investor confidence.

In that competition for attention and capital, the metaverse lost. Meta’s pullback was also not an isolated case. Microsoft moved away from metaverse-first ambitions as well, retiring the Immersive space (3D) view in Teams meetings, Microsoft Mesh on the web, and Mesh apps for PC and Quest in December 2025. The services were replaced by immersive events in Teams, a narrower offering built around specific workplace functions rather than a broad metaverse vision.

The wider retreat matters because it suggests the problem was not limited to Meta’s execution. Another major tech company also stepped back from standalone immersive environments and turned to more limited, use-specific tools instead. A larger pattern appeared from that shift: grand metaverse narratives gave way to practical features, embedded tools, and industry-specific uses. In that sense, the metaverse has not entirely disappeared, but it did lose its status as the next internet.

Metaverse’s afterlife in the age of AI

The metaverse’s decline does not necessarily imply a complete disappearance. What seems more likely is that parts of it will survive in altered form, detached from the sweeping vision that once surrounded it. Rather than continuing as a standalone digital world meant to transform social life, the metaverse may persist as a set of tools, features, and immersive functions folded into other technologies.

AI is likely to play a role in that transition. It can lower the cost of building virtual environments, speed up avatar creation, automate elements of interaction design, and make digital spaces more responsive. In this sense, AI may succeed where the original metaverse struggled, not by reviving the same vision, but by making parts of it more practical and easier to use.

Such a distinction is important because it shifts the focus from ideology to utility. The metaverse was once marketed as the next stage of the internet, yet its more durable applications now appear to lie in narrower settings where immersion serves a clear purpose. Training, design, simulation, and industrial planning are all contexts in which virtual environments can offer measurable value without becoming a universal social destination.

What might survive, then, is not the metaverse as it was originally imagined, but a smaller set of immersive capabilities embedded in gaming, education, industry, and workplace systems. Avatars, digital agents, simulations, and adaptive virtual spaces may all remain relevant, but as components rather than the foundation of a new social order.

The shift also helps explain the political lesson of the metaverse’s collapse. Large-scale investment, aggressive branding, and executive certainty were not enough to secure public legitimacy. Meta tried to present the metaverse as an inevitable horizon, yet users did not embrace it, markets did not reward it in proportion to the spending, and the company itself eventually narrowed the project it had once elevated into a corporate identity.

In that sense, the metaverse matters even in failure. Its retreat does not simply mark the end of an overhyped product cycle. It also reveals the limits of top-down corporate future-making, especially when private platforms try to define the direction of collective digital life before society has decided whether such a future is either desirable or necessary.

Conclusion

The metaverse failed because it asked too much of users, promised too much to investors, and concentrated too much power in a platform model that never convincingly earned public trust. Meta’s retreat from Horizon Worlds makes that failure difficult to ignore, while Microsoft’s parallel narrowing of immersive ambitions suggests the problem extended beyond one company’s misjudgement.

Immersive VR technologies are unlikely to vanish, and AI may even extend some of their useful applications. Yet the metaverse as a universal social future has largely collapsed under the combined weight of weak adoption, unsustainable economics, and an overly corporate vision of digital life. What remains is not the next internet, but a reminder that the future cannot simply be declared into existence by the companies most eager to own it.

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!

Edge AI advantages and challenges shaping the future of digital systems

Over the past few years, we have witnessed a rapid shift in the way data is stored and processed across businesses, organisations, and digital systems.

What we are increasingly seeing is that AI itself is changing form as computation shifts away from centralised cloud environments to the network edge. Such a shift has come to be known as edge AI.

Edge AI refers to the deployment of machine learning models directly on local devices such as smartphones, sensors, industrial machines, and autonomous systems.

Instead of transmitting data to remote servers for processing, analysis is performed on the device itself, enabling faster responses and greater control over sensitive information.

Such a transition marks a significant departure from earlier models of AI deployment, where cloud infrastructure dominated both processing and storage.

From centralised AI to edge intelligence

Traditional AI systems used to rely heavily on centralised architectures. Data collected from users or devices would be transmitted to large-scale data centres, where powerful servers would perform computations and generate outputs.

Such a model offered efficiency, scalability, and easier security management, as protection efforts could be concentrated within controlled environments.

Centralisation allowed organisations to enforce uniform security policies, deploy updates rapidly, and monitor threats from a single vantage point. However, reliance on cloud infrastructure also introduced latency, bandwidth constraints, and increased exposure of sensitive data during transmission.

Edge AI improves performance and privacy while expanding cybersecurity risks across distributed systems and devices.

Edge AI introduces a fundamentally different paradigm. Moving computation closer to the data source reduces the reliance on continuous connectivity and enables real-time decision-making.

Such decentralisation represents not merely a technical shift but a reconfiguration of the way digital systems operate and interact with their environments.

Advantages of edge AI

Reduced latency and real-time processing

Latency is significantly reduced when computation occurs locally. Edge systems are particularly valuable in time-sensitive applications such as autonomous vehicles, healthcare monitoring, and industrial automation, where delays can have critical consequences.

Enhanced privacy and data control

Privacy improves when sensitive data remains on-device instead of being transmitted across networks. Such an approach aligns with growing concerns around data protection, regulatory compliance, and user trust.

Operational resilience

Edge systems can continue functioning even when network connectivity is limited or unavailable. In remote environments or critical infrastructure, independence from central servers ensures service continuity.

Bandwidth efficiency and cost reduction

Bandwidth consumption is decreased because only processed insights are transmitted, not raw data. Such efficiency can translate into reduced operational costs and improved system performance.

Personalisation and context awareness

Devices can adapt to user behaviour in real time, learning from local data without exposing sensitive information externally. In healthcare, personalised diagnostics can be performed directly on wearable devices, while in manufacturing, predictive maintenance can occur on-site.

The dark side of edge AI

However, the shift towards edge computing introduces profound cybersecurity challenges. The most significant of these is the expansion of the attack surface.

Instead of a limited number of well-protected data centres, organisations must secure vast networks of distributed devices. Each endpoint represents a potential entry point for malicious actors.

The scale and diversity of edge deployments complicate efforts to maintain consistent security standards. Security is no longer centralised but dispersed, increasing the likelihood of vulnerabilities and misconfigurations.

Let’s take a closer look at some other challenges of edge AI.

Physical vulnerabilities and device exposure

Edge devices often operate in uncontrolled environments, making physical access a major risk. Attackers may tamper with hardware, extract sensitive information, or reverse engineer AI models.

hacker working computer with code

Model extraction attacks allow adversaries to replicate proprietary algorithms, undermining intellectual property and enabling further exploitation. Such risks are significantly more pronounced compared to cloud systems, where physical access is tightly controlled.

Software constraints and patch management challenges

Many edge devices rely on embedded systems with limited computational resources. Such constraints make it difficult to implement robust security measures, including advanced encryption and intrusion detection.

Patch management becomes increasingly complex in decentralised environments. Ensuring that millions of devices receive timely updates is a significant challenge, particularly when connectivity is inconsistent or when devices operate in remote locations.

Breakdown of traditional security models

The decentralised nature of edge AI undermines conventional perimeter-based security frameworks. Without a clearly defined boundary, traditional approaches to network defence lose effectiveness.

Each device must be treated as an independent security domain, requiring authentication, authorisation, and continuous monitoring. Identity management becomes more complex as the number of devices grows, increasing the risk of misconfiguration and unauthorised access.

Data integrity and adversarial threats

As we mentioned before, edge devices rely heavily on local data inputs to make decisions. As a result, manipulated inputs can lead to compromised outcomes. Adversarial attacks, in which inputs are deliberately altered to deceive machine learning models, represent a significant threat.

2910154 442

In safety-critical systems, such manipulation can lead to severe consequences. Altered sensor data in industrial environments may disrupt operations, while compromised vision systems in autonomous vehicles may produce dangerous behaviour.

Supply chain risks in edge AI

Edge AI systems depend on a combination of hardware, software, and pre-trained models sourced from multiple vendors. Each component introduces potential vulnerabilities.

Attackers may compromise supply chains by inserting backdoors during manufacturing, distributing malicious updates, or exploiting third-party software dependencies. The global nature of technology supply chains complicates efforts to ensure trust and accountability.

Energy constraints and security trade-offs

Edge devices are often designed with efficiency in mind, prioritising performance and power consumption. Security mechanisms such as encryption and continuous monitoring require computational resources that may be limited.

As a result, security features may be simplified or omitted, increasing exposure to cyber threats. Balancing efficiency with robust protection remains a persistent challenge.

Cyber-physical risks and real-world impact

The integration of edge AI into cyber-physical systems elevates the consequences of security breaches. Digital manipulation can directly influence physical outcomes, affecting safety and infrastructure.

Compromised healthcare devices may produce incorrect diagnoses, while disrupted transportation systems may lead to accidents. In energy networks, attacks could impact entire regions, highlighting the broader societal implications of edge AI vulnerabilities.

cybersecurity warning padlock red exclamation mark

Regulatory and governance challenges

Existing regulatory frameworks have been largely designed for centralised systems and do not fully address the complexities of decentralised architectures. Questions regarding liability, accountability, and enforcement remain unresolved.

Organisations may struggle to implement effective security practices without clear standards. Policymakers face the challenge of developing regulations that reflect the distributed nature of edge AI systems.

Towards a secure edge AI ecosystem

Addressing all these challenges requires a multi-layered and adaptive approach that reflects the complexity of edge AI environments.

Hardware-level protections, such as secure enclaves and trusted execution environments, play a critical role in safeguarding sensitive operations from physical tampering and low-level attacks.

Encryption and secure boot processes further strengthen device integrity, ensuring that both data and models remain protected and that unauthorised modifications are prevented from the outset.

At the software level, continuous monitoring and anomaly detection are essential for identifying threats in real time, particularly in distributed systems where central oversight is limited.

Secure update mechanisms must also be prioritised, ensuring that patches and security improvements can be deployed efficiently and reliably across large networks of devices, even in conditions of intermittent connectivity.

Without such mechanisms, vulnerabilities can persist and spread across the ecosystem.

data breach laptop exploding cyber attack concept

At the same time, many enterprises are increasingly adopting a hybrid approach that combines edge and cloud capabilities.

Rather than relying entirely on decentralised or centralised models, organisations are distributing workloads strategically, keeping latency-sensitive and privacy-critical processes on the edge while maintaining centralised oversight, analytics, and security coordination in the cloud.

Such an approach allows organisations to balance performance and control, while enabling more effective threat detection and response through aggregated intelligence.

Security must also be embedded into system design from the outset, rather than treated as an additional layer to be applied after deployment. A proactive approach to risk assessment, combined with secure development practices, can significantly reduce vulnerabilities before systems are operational.

Furthermore, collaboration between industry, governments, and research institutions will be crucial in establishing common standards, improving interoperability, and ensuring that security practices evolve alongside technological advancements.

In conclusion, we have seen how the rise of edge AI represents a pivotal shift in both AI and cybersecurity. Decentralisation enables faster, more private, and more resilient systems, yet it also creates a fragmented and dynamic attack surface.

The advantages we have outlined are compelling, but they also introduce additional layers of complexity and risk. Addressing these challenges requires a comprehensive approach that combines technological innovation, regulatory development, and organisational awareness.

Only through such coordinated efforts can the benefits of edge AI be realised while ensuring that security, trust, and safety remain intact in an increasingly decentralised digital landscape.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

Advancing global digital cooperation and AI innovation across the UN system

Digital technologies and AI are increasingly shaping economic development, governance and international cooperation. As these technologies expand rapidly, international organisations are working to ensure that innovation is accompanied by responsible governance, inclusive access and coordinated global policies.

Within the United Nations system, a range of initiatives aim to strengthen cooperation on digital transformation and the development of AI. These efforts address issues such as digital infrastructure, data governance, technological innovation and equitable participation in emerging digital ecosystems. International collaboration plays an essential role in ensuring that the benefits of digital technologies support sustainable development while reducing global inequalities in access to digital resources.

Several programmes across the United Nations system reflect these priorities, combining global governance initiatives with practical AI applications in areas such as development, humanitarian response and digital inclusion. The following sections examine selected initiatives that illustrate how AI and digital cooperation are being advanced across different areas of the UN system.

Global Digital Compact

City

The Global Digital Compact is a comprehensive international framework adopted by United Nations member states to guide global digital cooperation and enhance the governance of AI. Negotiated by the 193 member states and reflects broad consultations aimed at shaping a shared vision for a digital future that is open, inclusive, safe, and secure for all. The Compact is part of the Pact for the Future, adopted at the 2024 Summit of the Future in New York.

At its core, the Compact seeks to address persistent digital divides by promoting universal connectivity, affordable access and inclusive participation in the digital economy. Governments and stakeholders have committed to connecting all individuals, schools, and hospitals to the internet, increasing investment in digital public infrastructure, and ensuring that technologies are accessible in diverse languages and formats.

The Compact also emphasises human rights and the protection of fundamental freedoms in the digital space, calling for the strengthened legal and policy frameworks that uphold international law and protect users from harms such as misinformation and discrimination. It promotes an open, global, stable, and secure internet while supporting access to independent, fact-based information.

The key objective of the Compact is to enhance international cooperation on data governance and AI for the benefit of humanity. It includes commitments to develop interoperable national data governance frameworks, advance responsible and equitable approaches to AI governance, and establish mechanisms for global dialogue and scientific guidance on AI. These elements reflect the need for collaborative, multistakeholder governance that balances innovation with transparency, accountability, and respect for human rights.

Independent International Scientific Panel on AI

Logo, Text

The Independent International Scientific Panel on AI is a mechanism called for within the Global Digital Compact to support evidence‑based policymaking in AI governance. Member states requested the establishment of a multi‑disciplinary panel under the United Nations to assess the opportunities, risks and societal impacts of AI, and to promote scientific understanding across geographic and sectoral divides.

The panel is intended to contribute robust, independent scientific analysis to global AI discussions, ensuring that policy decisions are grounded in research rather than short‑term market pressures or fragmented national approaches. Its mandate includes conducting comprehensive risk and impact assessments, developing common methodologies for evaluating AI systems, and advising on interoperable governance frameworks that respect human rights and international law.

By bringing together experts from diverse disciplines and regions, the panel aims to bridge the gap between scientific developments and policymaking. It is a key institutional mechanism for fostering inclusive AI governance, with balanced geographic representation to ensure that insights reflect global needs rather than narrow technological interests.

The panel also complements the broader Global Dialogue on AI Governance, which seeks to engage governments, international organisations, civil society and technical communities in ongoing discussions about normative approaches, standards, and principles for global AI governance.

The UN Digital Cooperation Portal

The UN Digital Cooperation Portal is a central platform designed to support the implementation of the Global Digital Compact by mapping global digital cooperation activities and facilitating coordination among diverse stakeholders. The portal invites governments, UN entities, civil society organisations, researchers, and private sector actors to voluntarily submit information on initiatives related to the Compact’s objectives.

Launched in December 2025, the portal aggregates initiatives across thematic areas, including digital inclusion, AI governance, data governance, digital infrastructure, and the protection of human rights online. By visualising how activities align with agreed international frameworks, the platform supports strategic collaboration, strengthens transparency and highlights opportunities for joint action across regions and sectors.

The portal generates interactive data visualisations that illustrate how digital cooperation initiatives are evolving at the national, regional and global levels. These tools help identify gaps and overlaps in current efforts, enabling stakeholders to coordinate more effectively in pursuit of shared objectives such as closing digital divides and advancing equitable digital development.

As a resource for governments, UN agencies and external partners, the portal also contributes to the preparatory process for the high‑level review of the Global Digital Compact scheduled for 2027, providing an evidence‑based foundation assessing progress and emerging policy priorities.

Closing the language gap in AI through local language accelerators

Text, Symbol

Language diversity remains one of the major challenges in global AI development. More than half of the world’s population speaks one of over seven thousand languages, yet most AI systems currently support only a small number of widely used global languages.

Around 1.2 billion people rely on low-resource languages that remain poorly represented in digital technologies. Limited language representation can restrict access to AI-powered services in sectors such as agriculture, healthcare, education and civic participation.

The Local Language Accelerators programme, developed by the United Nations Development Programme, addresses this challenge by supporting the creation of digital language resources and AI models for underrepresented languages.

The initiative combines technological development with partnerships involving universities, research institutions and local language communities. The technologies involved include optical character recognition systems that digitise written texts, automatic speech recognition tools capable of processing spoken language and text-to-speech technologies that generate digital audio.

Ten projects are currently underway across four continents, including initiatives in Serbia, the Democratic Republic of the Congo, the Republic of the Congo, Namibia, Lesotho, Ghana, Mexico, Peru, Nepal and Iraq. These projects support the creation of new datasets and language resources that can be reused for future AI systems.

Using satellite imagery and AI to improve disaster response

Logo, Text

Rapid damage assessment plays a critical role in humanitarian response following natural disasters. Traditional assessment methods often require manual analysis of satellite images and field inspections conducted by experts, a process that can take weeks.

Emergency response operations, however, require reliable information within the first seventy-two hours after a disaster to prioritise rescue operations and humanitarian assistance.

The SKAI platform, developed by the World Food Programme Innovation Accelerator, uses AI-based computer vision to analyse satellite imagery and identify damaged buildings automatically. The system enables humanitarian organisations to assess destruction at the level of individual structures across large geographic areas.

Developed as an open-source project in collaboration with Google Research, the platform can generate prioritised damage assessments within approximately twenty-four hours. Since 2022, the system has analysed more than 3.9 million buildings and identified around 450,000 severely damaged or destroyed structures.

Expanding inclusive participation through the UN Women AI School

Logo, Text, Outdoors

Increasing participation in AI development is another priority across the United Nations system. Women remain underrepresented in many AI-related fields, including machine learning engineering and data science.

The UN Women AI School addresses this challenge by providing training programmes designed for policymakers, civil society organisations, UN staff, and young innovators. The initiative aims to strengthen AI literacy and encourage broader participation in shaping the future of digital technologies.

Participants follow structured training tracks combining technical education with discussions on AI governance, ethics, and social impact. Collaborative learning environments encourage participants to develop solutions tailored to the needs of their communities.

More than three thousand participants have taken part in the programme since its launch. A train-the-trainer (ToT) model enables graduates to support future training programmes and expand the initiative to additional regions.

Responsible AI in satellite technologies and earth observation

Logo, Outdoors

AI technologies are increasingly integrated into satellite systems and Earth observation platforms. These systems analyse large volumes of geospatial data and generate near-real-time insights about environmental conditions.

Applications include monitoring climate change, analysing natural disasters, and supporting environmental policy planning. Rapid technological progress in this field also raises governance challenges related to transparency and accountability.

Many AI models used in satellite analysis operate as black box systems whose internal decision-making processes are difficult to interpret. Limited transparency can create risks when such systems are used to inform critical policy decisions.

Data bias represents another concern. Training datasets often originate primarily from the Global North, which may lead to inaccurate interpretations of environmental conditions in other regions of the world.

Experts from the United Nations Office for Outer Space Affairs have therefore proposed a framework promoting the responsible use of AI in space technologies. The framework emphasises transparency, accountability, and continued human oversight.

Assessing national readiness for AI governance

License Plate, Transportation, Vehicle

UNESCO’s AI Readiness Assessment Methodology helps governments evaluate their capacity to adopt and regulate AI technologies responsibly.

The methodology examines multiple dimensions of national AI ecosystems, including infrastructure, research capacity, institutional readiness and regulatory frameworks. Rather than ranking countries, the assessment identifies strengths and areas requiring further development.

Since its introduction in 2022, the methodology has been implemented in more than seventy countries. More than seventeen thousand stakeholders have participated in consultations associated with the initiative.

Assessment results have contributed to the development of national AI strategies and policy frameworks in several regions. An updated version of the methodology is expected to be released in 2026.

Additionally, UNESCO promotes the ethical development and use of AI through its Recommendation on the Ethics of Artificial Intelligence. The global framework sets out principles on transparency, accountability, fairness, and respect for human rights to guide national policies and international cooperation.

AI for Good and global capacity building

Art, Logo

The International Telecommunication Union coordinates the AI for Good initiative, which focuses on applying AI technologies to global challenges while strengthening international cooperation in governance and standards.

The programme operates across multiple areas, including multistakeholder dialogue, technical standard development, governance support and capacity development activities.

More than four hundred AI-related standards have already been developed in areas such as multimedia technologies, energy efficiency and cybersecurity. Governance dialogues organised through the initiative have involved more than one hundred ministers and regulators.

Educational programmes linked to the initiative aim to expand digital skills among young people worldwide through robotics competitions, machine learning challenges and educational partnerships.

The AI for Good Global Summit 2026, set to take place from 7–10 July in Geneva, will convene governments, industry leaders and civil society to advance AI governance, promote responsible innovation, and highlight initiatives that foster inclusive and equitable digital development.

AI tools supporting refugee entrepreneurship

Logo

AI technologies are also being used to support the economic opportunities for displaced populations. The United Nations Refugee Agency has developed an AI-powered virtual assistant designed to help refugees and asylum seekers transform business ideas into structured business plans.

The platform guides users through financial planning, market analysis and the preparation of investment proposals. The development of the system involved collaboration with NGOs, governments, and entrepreneurial networks across Latin America.

The tool was initially implemented in Paraguay and was designed with input from refugee communities. Remote access allows users to engage with the platform regardless of geographical or institutional constraints.

More than 340 refugee entrepreneurs have used the platform since its launch, with women representing approximately sixty percent of participants. The model is designed to be scalable and could be implemented in additional regions.

Promoting responsible innovation in civilian AI for peace and security

Logo

The rapid expansion of AI technologies brings increasing security challenges, particularly due to the potential misuse of civilian AI systems in military, conflict-related, or high-risk contexts. Dual-use applications mean that tools designed for civilian purposes, such as data analysis or autonomous systems, could also be repurposed in ways that threaten international peace, stability or human safety.

The United Nations Office for Disarmament Affairs works to foster responsible innovation practices, ensuring that the development and deployment of AI technologies consider their broader implications for global peace and security. Addressing these risks requires ongoing collaboration and dialogue among policymakers, researchers, industry stakeholders, and civil society, creating a shared framework for understanding and mitigating potential threats.

To support this, the programme organises a comprehensive set of initiatives, including thematic multistakeholder dialogues, academic workshops, public panels, private sector roundtables and in-person training sessions for graduate students. These activities aim not only to raise awareness of emerging security risks, but also to provide practical guidance and tools that promote safe, transparent and accountable AI practices in civilian applications worldwide.

UN 2.0 Communities of Practice

Advertisement, Poster, Text, QR Code, Person, Head

Knowledge sharing and collaboration are strengthened through UN 2.0 Communities of Practice, connecting partners across the United Nations system and beyond. The networks facilitate the exchange of expertise and approaches on digital transformation, data strategy, innovation, and strategic foresight.

Over 18,000 practitioners from more than 160 countries participate, enhancing the collective capacity to address complex AI and digital challenges. Thematic groups, including those focused on digital and data initiatives, support peer-to-peer engagement, professional development, and collaborative problem-solving. Participation allows stakeholders to contribute to a wider ecosystem of expertise and innovation, promoting inclusive digital governance and supporting the Sustainable Development Goals.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot! 

Anthropic’s Pentagon dispute and military AI governance in 2026

On 28 February 2026, Anthropic’s Claude rose to No. 1 in Apple’s US App Store free rankings, overtaking OpenAI’s ChatGPT. The surge came shortly after OpenAI announced a partnership with the US Department of Defense (DoD), making its technology available to the US Army. The development prompted discussion among users and observers about whether concerns over military partnerships were influencing the shift to alternative AI tools.

Mere hours before the USD $200 million OpenAI-DoD deal was finalised, Anthropic was informed that its potential deal with the Pentagon had fallen through, largely because the AI company refused to relinquish total control of its technology for domestic mass surveillance. According to reporting, discussions broke down after Anthropic declined to grant the US government unrestricted control over its models, particularly for potential uses related to large-scale surveillance.

Following the breakdown of negotiations, US officials reportedly designated Anthropic as a ‘supply chain risk to national security’. The decision effectively limited the company’s participation in certain defence-related projects and highlighted growing tensions between AI developers’ safety policies and government expectations regarding national security technologies.

The debate over military partnerships sparked internal and industry-wide discussion. Caitlin Kalinowski, the former head of AR glasses hardware at Meta and the hardware leader at OpenAI, resigned soon after the US DoD deal, citing ethical concerns about the company’s involvement in military AI applications.

AI has driven recent technological innovation, with companies like Anduril and Palantir collaborating with the US DoD to deploy AI on and off the battlefield. The debate over AI’s role in military operations, surveillance, and security has intensified, especially as Middle East conflicts highlight its potential uses and risks.

Against this backdrop, the dispute between Anthropic and the Pentagon reflects a wider debate on how AI should be used in security and defence. Governments are increasingly relying on private tech companies to develop the systems that shape modern military capabilities, while those same companies are trying to set limits on how their technologies can be used.

As AI becomes more deeply integrated into security strategies around the world, the challenge may no longer be whether the technology will be used, but how it should be governed. The question is: who should ultimately decide where the limits of military AI lie?

Anthropic’s approach to military AI

Anthropic’s approach is closely tied to its concept of ‘constitutional AI’, a training method that guides how the model behaves by embedding a set of principles directly into its responses. Such principles are intended to reduce harmful outputs and ensure the system avoids unsafe or unethical uses. While such safeguards are intended to improve reliability and trust, they can also limit how the technology can be deployed in more sensitive contexts such as military operations.

Anthropic’s Constitution says its AI assistant should be ‘genuinely helpful’ to people and society, while avoiding unsafe, unethical, or deceptive actions. The document reflects the company’s broader effort to build safeguards into model deployment. In practice, Anthropic has set limits on certain applications of its technology, including uses related to large-scale surveillance or military operations.

Anthropic presents these safeguards as proof of its commitment to responsible AI. Reports indicate that concerns over unrestricted model access led to the breakdown in talks with the US DoD.

At the same time, Anthropic clarifies that its concerns are specific to certain uses of its technology. The company does not generally oppose cooperation with national security institutions. In a statement following the Pentagon’s designation of the company as a ‘supply chain risk to national security’, CEO Dario Amodei said, ‘Anthropic has much more in common with the US DoD than we have differences.’ He added that the company remains committed to ‘advancing US national security and defending the American people.’

The episode, therefore, highlights a nuanced position. Anthropic appears open to defence partnerships but seeks to maintain clearer limits on the deployment of its AI systems. The disagreement with the Pentagon ultimately reflects not a fundamental difference in goals, but rather different views on how far military institutions should be able to control and use advanced AI technologies.

Anthropic’s position illustrates a broader challenge facing governments and tech companies as AI becomes increasingly integrated into national security systems. While military and security institutions are eager to deploy advanced AI tools to support intelligence analysis, logistics, and operational planning, the companies developing these technologies are also seeking to establish safeguards for their use. Anthropic’s willingness to step back from a major defence partnership and challenge the Pentagon’s response underscores how some AI developers are trying to set limits on military uses of their systems.

Defence partnerships that shape the AI industry

While Anthropic has taken a cautious approach to military deployment of AI, other technology companies have pursued closer partnerships with defence institutions. One notable example is Palantir, the US data analytics firm co-founded by Peter Thiel that has longstanding relationships with numerous government agencies. Documents leaked in 2013 suggested that the company had contracts with at least 12 US government bodies. More recently, Palantir has expanded its defence offering through its Artificial Intelligence Platform (AIP), designed to support intelligence analysis and operational decision-making for military and security institutions.

Another prominent player is Anduril Industries, a US defence technology company focused on developing AI-enabled defence systems. The firm produces autonomous and semi-autonomous technologies, including unmanned aerial systems and surveillance platforms, which it supplies to the US DoD.

Shield AI, meanwhile, is developing autonomous flight software designed to operate in environments where GPS and communications may be unavailable. Its Hivemind AI platform powers drones that can navigate buildings and complex environments without human control. The company has worked with the US military to test these systems in training exercises and operational scenarios, including aircraft autonomy projects aimed at supporting fighter pilots.

The aforementioned partnerships illustrate how the US government has increasingly embraced AI as a key pillar of national defence and future military operations. In many cases, these technologies are already being used in operational contexts. Palantir’s Gotham and AIP, for instance, have supported US military and intelligence operations by processing satellite imagery, drone footage, and intercepted communications to help analysts identify patterns and potential threats.

Other companies are contributing to defence capabilities through autonomous systems development and hardware integration. Anduril supplies the US DoD with AI-enabled surveillance, drone, and counter-air systems designed to detect and respond to potential threats. At the same time, OpenAI’s technology is increasingly being integrated into national security and defence projects through growing collaboration with US defence institutions.

Such developments show that AI is no longer a supporting tool but a fundamental part of military infrastructure, influencing how defence organisations process information and make decisions. As governments deepen their reliance on private-sector AI, the emerging interplay among innovation, operational effectiveness, and oversight will define the central debate on military AI adoption.

The potential benefits of military AI

The debate over Anthropic’s restrictions on military AI use highlights the reasons governments invest in such technologies: defence institutions are drawn to AI because it processes vast amounts of information much faster than human analysts. Military operations generate massive data streams from satellites, drones, sensors, and communication networks, and AI systems can analyse them in near real time.

In 2017, the US DoD launched Project Maven to apply machine learning to drone and satellite imagery, enabling analysts to identify objects, movements, and potential threats on the battlefield faster than with traditional manual methods.

AI is increasingly used in military logistics and operational planning. It helps commanders anticipate equipment failures, enables predictive maintenance, optimises supply chains, and improves field asset readiness.

Recent conflicts have shown that AI-driven tools can enhance military intelligence and planning. In Ukraine, for example, forces reportedly used software to analyse satellite imagery, drone footage, and battlefield data. Key benefits include more efficient target identification, real-time tracking of troop movements, and clearer battlefield awareness through the integration of multiple data sources.

AI-assisted analysis has been used in intelligence and targeting during the Gaza conflict. Israeli defence systems use AI tools to rapidly process large datasets for surveillance and intelligence operations. The tools help analysts identify potential militant infrastructure, track movements, and prioritise key intelligence, thus speeding up information processing for teams during periods of high operational activity.

More broadly, AI is transforming the way militaries coordinate across land, air, sea, and cyber domains. AI integrates data from diverse sources, equipping commanders to interpret complex operational situations and enabling faster, informed decision-making. The advances reinforce why many governments see AI as essential for future defence planning.

Ethical concerns and Anthropic’s limits on military AI

Despite the operational advantages of military AI, its growing role in national defence systems has raised ethical concerns. Critics warn that overreliance on AI for intelligence analysis, targeting, or operational planning could introduce risks if the systems produce inaccurate outputs or are deployed without sufficient human oversight. Even highly capable models can generate misleading or incomplete information, which in high-stakes military contexts could have serious consequences.

Concerns about the reliability of AI systems are also linked to the quality of the data they learn from. Many models still struggle to distinguish authentic information from synthetic or manipulated content online. As generative AI becomes more widespread, the risk that systems may absorb inaccurate or fabricated data increases, potentially affecting how these tools interpret intelligence or analyse complex operational environments.

Questions about autonomy have also become a major issue in discussions around military AI. As AI systems become increasingly capable of analysing battlefield data and identifying potential targets, debates have emerged over how much decision-making authority they should be given. Many experts argue that decisions involving the use of lethal force should remain under meaningful human control to prevent unintended consequences or misidentification of targets.

Another area of concern relates to the potential expansion of surveillance capabilities. AI systems can analyse satellite imagery, communications data, and online activity at a scale beyond the capacity of human analysts alone. While such tools may help intelligence agencies detect threats more efficiently, critics warn that they could also enable large-scale monitoring if deployed without clear legal and institutional safeguards.

It is within this ethical landscape that Anthropic has attempted to position itself as a more cautious actor in the AI industry. Through initiatives such as Claude’s Constitution and its broader emphasis on AI safety, the company argues that powerful AI systems should include safeguards that limit harmful or unethical uses. Anthropic’s reported refusal to grant the Pentagon unrestricted control over its models during negotiations reflects this approach.

The disagreement between Anthropic and the US DoD therefore highlights a broader tension in the development of military AI. Governments increasingly view AI as a strategic technology capable of strengthening defence and intelligence capabilities, while some developers seek to impose limits on how their systems are deployed. As AI becomes more deeply embedded in national security strategies, the question may no longer be whether these technologies will be used, but who should define the boundaries of their use.

Military AI and the limits of corporate control

Anthropic’s dispute with the Pentagon shows that the debate over military AI is no longer only about technological capability. Questions of speed, efficiency, and battlefield advantage now collide with concerns over surveillance, autonomy, human oversight, and corporate responsibility. Governments increasingly see AI as a strategic asset, while companies such as Anthropic are trying to draw boundaries around how far their systems can go once they enter defence environments.

Contrasting approaches across the industry make the tension even clearer. Palantir, Anduril, Shield AI, and OpenAI have moved closer to defence partnerships, reflecting a broader push to integrate advanced AI into military infrastructure. Anthropic, by comparison, has tried to keep one foot in national security cooperation while resisting uses it views as unsafe or unethical. A divide of that kind suggests that the future of military AI may be shaped as much by company policies as by government strategy.

The growing reliance on private firms to build national security technologies has made governance harder to define. Military institutions want flexibility, scale, and operational control, while AI developers increasingly face pressure to decide whether they are simply suppliers or active gatekeepers of how their models are deployed. Anthropic’s position does not outright defence cooperation, but it does expose how fragile the relationship becomes when state priorities and corporate safeguards no longer align.

Military AI will continue to expand, whether through intelligence analysis, logistics, surveillance, or autonomous systems. Governance, however, remains the unresolved issue at the centre of that expansion. As AI becomes more deeply embedded in defence policy and military planning, should governments alone decide how far these systems can go, or should companies like Anthropic retain the power to set limits on their use?

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!

Diplo chatbot can make mistakes. Consider checking important information.